Umi Save Image (with metadata)

Here's a problem that only shows up once you start using wildcards seriously. You queue sixty variations, get one you love, and the PNG's embedded metadata tells you the prompt was __hair__ girl in a __location__, {red|blue} jacket - the template, not the roll. Nothing in the file tells you which hair color actually came out. Umi Save Image exists to close that gap: it writes the resolved prompt into the PNG alongside the original template, so the picture is still searchable and reproducible after the fact.

What it does

It's a save node in the ordinary ComfyUI sense - an output node that writes PNGs into ComfyUI/output/ - with one difference: it takes the prompt as wiring rather than reading it out of the graph. Two required inputs, images (the batch to save) and filename_prefix (default Umi, and it accepts a subfolder, e.g. portraits/Umi; a counter is appended automatically, and %batch_num% addresses the batch index).

Then four optional inputs, all of which are connection-only sockets - you cannot type into them:

• positive_prompt ← the Wildcard Processor's text output (the resolved prompt)
• negative_prompt ← negative_text
• input_prompt ← input_text (the original template, before expansion)
• input_negative ← input_negative

Wire all four and each PNG carries umi_prompt, umi_negative, umi_input_prompt and umi_input_negative as text chunks, on top of the usual ComfyUI workflow and prompt chunks. That's what the pack's Image Browser panel scans when it searches output images - it reads exactly those umi_ keys, falling back to the stock prompt chunk.

Why the template matters

Most people only wire the first two. Don't skip input_text. The resolved prompt tells you what generated this image; the template tells you what to change to get a different one. Without it you're reverse-engineering your own prompt language from one lucky sentence - which, if you've ever tried to reuse a wildcard batch a month later, you know is a miserable afternoon.

Install

It's part of the pack core, so there's no separate install:

cd ComfyUI/custom_nodes
git clone https://github.com/Tinuva88/Comfy-UmiAI
cd Comfy-UmiAI && pip install -r requirements.txt

Restart ComfyUI afterwards. Only pyyaml is needed for core; the image browser that reads this metadata back is in the optional UI-tools overlay, which also wants requests and curl_cffi.

Where people get burned

Nothing connected, nothing written. Because those four sockets are connection-only, an unwired node silently produces a normal PNG with no Umi metadata, and the browser then shows nothing for it - which reads as "the metadata feature is broken." The node prints a warning to the console in that case, so check the console before filing a bug. The same happens if ComfyUI was launched with --disable-metadata: the node says so and saves the image anyway.

This is not the CivitAI node. Umi Save Image writes ComfyUI-style graph chunks plus the four umi_ fields. It does not write the flat A1111-style parameters block, which is what CivitAI's uploader parses to auto-link your checkpoint and LoRAs. If auto-linking on upload matters to you, that's a different save node's job - this one is for your own workflow and your own browser.

PNG only. The metadata lives in PNG text chunks. Convert the file to JPEG or WebP, or let a host re-encode it, and it's gone - same rule as any ComfyUI output. Keep the PNG, publish the small version.

One last thing: an empty image batch doesn't error. It logs that it received nothing and saves nothing, so if a save node appears to do nothing at all, look upstream at whatever produced zero images.